Speech recognition in noise using a self-structuring noise reduction model and hidden control models

Abstract

The author describes how speech recognition in the presence of F-16 jet cockpit noise can be performed using a sequence of three units, i.e. an auditory model and two neural models. A method for noise reduction in the cepstral domain based on a self-structuring universal approximator is proposed and tested on a large database of isolated words contaminated with jet noise. This approach is a potential alternative to traditional recognition methods for noisy speech and makes noise reduction possible in all three models. The first model performs a spectral analysis of the input speech signal. The second model is a self-structuring neural noise reduction (SNNR) model, which is an alternative to the noise reduction model. The noise-reduced output from the SNNR network is propagated through the speech recognizer consisting of a set of hidden control neural networks (HCNN). The author concludes that the SNNR network is a very powerful method for noise reduction in general and that the preliminary results presented can be improved. >